B-mode超音波腹部影像多器官識別研究：假體實驗

Identification of Abdominal Organs for B-mode Ultrasound Images Using Deep Learning Methods：A Phantom Study

黃勤凱(Cin-Kai Huang)

指導教授 ： 陳泰賓

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

目的：本研究應用深度學習網路方法自動分割超音波影像中的腹部器官，使用不同訓練參數及分割模型觀察腹部超音波影像識別並進行實驗結果評估。
方法與材料：採用監督式學習，參考卷積神經網路包含Mobilenetv2、Resnet18及ResNet50架構，通過轉移學習方式進行訓練，從中查找這些模型中合適的訓練比例、優化器、批量大小及訓練時期。藉由超音波儀器進行掃描超音波人體組織假體，掃瞄範圍為腹部主要於影像中可觀察到肝臟、膽囊、胰臟、脾臟和腎臟之位置，影像收集研判由各種不同角度與深度獲取。從電腦中擷取影像，原始影像為720×480像素尺寸，總共收集600張影像。
結果：分割性能評估包括整體準確率、平均準確率、平均交疊率、加權交疊率及平均邊界F1分數。本研究可初步了解全卷積神經網路對假體影像識別之成效，ResNet50分割模型在測試集之效能在肝臟、胰臟、右腎以及左腎中具有最好的表現，其中這5項指標結果分別為0.983、0.971、0.955、0.967、0.883；0.994、0.862、0.807、0.988、0.877；0.992、0.960、0.919、0.985、0.860；0.991、0.959、0.902、0.984、0.856。Mobilenetv2分割模型在測試集之效能在膽囊及脾臟中具有最好的表現，其中這5項指標結果分別為0.994、0.898、864、0.988、0.838；0.987、0.892、0.869、0.975、0.806。
結論：根據實驗結果，所提出的方法已經證明重疊率可以達到八成以上分割效能。此外，所提出的方法可以通過使用或調查未來最新的基於卷積神經網路的模型來提高腹部多器官的分割準確性。

關鍵字

腹部超音波 ； 深度學習 ； 全卷積網路

並列摘要

Purpose: This study automatically segments abdominal organs in ultrasound images, applying a deep learning network approach. Experimental evaluation shows the identification of abdominal ultrasound images through different parameters and segmentation models.
Methods and Materials: Meanwhile, using supervised learning, refer to the Mobilenetv2, Resnet18, and ResNet50 architecture of the convolutional neural network for training by transfer learning. A sequential searching technique was applied to find the training ratio, optimizer, batch size, and numbers of epochs in these models. Scan the ultrasonic tissue phantom with the ultrasonic instrument. The locations of imaging were along with the epigastric region. The organs included the liver, gallbladder, pancreas, spleen, and kidney. Imagery exploitations were acquired from various angles and depths, the original image was 720×480 pixels in size, and a total of 600 images were collected.
Results: The evaluation of segmented performance included global accuracy, mean accuracy, mean intersection over union (IOU), weighted IoU, and mean boundary F1 Score (BF Score). This study has an elementary comprehension of the effect of identifying phantom images by the fully convolutional neural network. In the results, the Resnet50 has the high performance in segmentation among liver, pancreas, right kidney, and left kidney. The global accuracy, mean accuracy, mean IoU, weighted IoU, and mean BF score as 0.983, 0.971, 0.955, 0.967, and 0.883; 0.994, 0.862, 0.807, 0.988, and 0.877; 0.992, 0.960, 0.919, 0.985, and 0.860; 0.991, 0.959, 0.902, 0.984, and 0.856. In the results, the Mobilenetv2 has the high performance in segmentation both gallbladder and spleen. The global accuracy, mean accuracy, mean IoU, weighted IoU, and mean BF score as 0.994, 0.898, 0.864, 0.988, and 0.987; 0.892, 0.869, 0.975, 0.806.
Conclusion: According to the experimental results, the presented methods have proven the overlap rate over 80% of the segmentation efficiency. Furthermore, the presented methods could increase the accuracy of segmentation for multiple organs in the abdomen by using or surveying the newest CNN-based models in the future.

並列關鍵字

Abdominal Ultrasound ； Deep Learning ； FCN

參考文獻

1. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi F, Ghafoorian M, van der Laak JAWM, van Ginneken B, Sánchez CI. A survey on deep learning in medical image analysis. Med Image Anal. 2017 Dec; 42:60-88. doi: 10.1016/j.media.2017.07.005. Epub 2017 Jul 26. PMID: 28778026.

2. Komura D, Ishikawa S. Machine learning approaches for pathologic diagnosis. Virchows Arch. 2019 Aug;475(2):131-138. doi: 10.1007/s00428-019-02594-w. Epub 2019 Jun 20. PMID: 31222375.

3. Tomizawa M, Shinozaki F, Hasegawa R, Shirai Y, Motoyoshi Y, Sugiyama T, Yamamoto S, Ishige N. Abdominal ultrasonography for patients with abdominal pain as a first-line diagnostic imaging modality. Exp Ther Med. 2017 May;13(5):1932-1936. doi: 10.3892/etm.2017.4209. Epub 2017 Mar 9. PMID: 28565789.

4. Xu Y, Jia Z, Wang LB, Ai Y, Zhang F, Lai M, Chang EI. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC Bioinformatics. 2017 May 26;18(1):281. doi: 10.1186/s12859-017-1685-x. PMID: 28549410.

5. Seo J, Kim YS. Ultrasound imaging and beyond: recent advances in medical ultrasound. Biomed Eng Lett. 2017 Apr 14;7(2):57-58. doi: 10.1007/s13534-017-0030-7. PMID: 30603151.